Machine-learning graph convolutional electronic propagators.

Area of Science:

• Quantum mechanics
• Computational chemistry
• Machine learning

Background:

• Simulating time evolution of quantum systems is computationally intensive.
• Existing methods struggle with scalability for complex molecular and condensed-phase systems.

Purpose of the Study:

• To develop a novel graph-based machine learning framework for simulating electronic dynamics.
• To introduce and evaluate two model variants: one for wavefunctions and one for electron densities.

Main Methods:

• Utilized a recursive Chebyshev graph neural network architecture.
• Trained models on trajectory data from tight-binding and electron-phonon coupled systems.
• Investigated both complex-valued wavefunction and electron density propagation.

Main Results:

• Wavefunction-based models achieved near-exact long-time propagation in various regimes.
• Density-only models showed strong performance with physics-informed loss functions.
• Demonstrated the potential for resolution-independent electronic dynamics simulation.

Conclusions:

• The graph-based framework provides a foundation for scalable quantum simulations.
• This approach opens new pathways for studying complex quantum systems.
• The developed models offer efficient and accurate simulation of electronic processes.